关键词: 模态分解, 深度信念网络, 集合卡尔曼滤波, 径流预测, 组合模型

Abstract: Accurate short-term runoff prediction can provide important scientific basis for water resources planning, flood control and drought relief in river basin. To mitigate systematic errors and enhance runoff prediction accuracy of models, we propose the decomposition-correction-integration framework based on the decomposition-integration model. Within this framework, we construct the EEMD-DBN-EnKF and VMD-DBN-EnKF models. Leveraging the Ensemble Kalman Filter data assimilation algorithm, we correct components deviating significantly from measured runoff to alleviate systematic errors introduced by the decomposition process in prediction. Comparative analysis is conducted against the unmodified EEMD-DBN, VMD-DBN, and single DBN models. Results demonstrate that the combination model based on modal decomposition reduces RMSE by a minimum of 23% compared to individual models, while NSE and R² increase by over 21%. Notably, the runoff component-corrected combined model exhibits improved evaluation metrics relative to the modal decomposition-based model. Among these models, the VMD-DBN-EnKF prediction model exhibits the least error and highest effectiveness, with NSE and R² exceeding 0.89, followed by EEMD-DBN-EnKF, VMD-DBN, and EEMD-DBN in descending order. In conclusion, the “decomposition-correction-integration” prediction framework demonstrates robust applicability in the Manas River Basin, offering valuable technical support for short-term runoff forecasts.

Key words: modal decomposition, deep belief network, ensemble Kalman filter, runoff prediction, combination model